mm/huge_memory.c

   1 /*
   2  *  Copyright (C) 2009  Red Hat, Inc.
   3  *
   4  *  This work is licensed under the terms of the GNU GPL, version 2. See
   5  *  the COPYING file in the top-level directory.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/highmem.h>
  11 #include <linux/hugetlb.h>
  12 #include <linux/mmu_notifier.h>
  13 #include <linux/rmap.h>
  14 #include <linux/swap.h>
  15 #include <linux/mm_inline.h>
  16 #include <linux/kthread.h>
  17 #include <linux/khugepaged.h>
  18 #include <linux/freezer.h>
  19 #include <linux/mman.h>
  20 #include <linux/pagemap.h>
  21 #include <asm/tlb.h>
  22 #include <asm/pgalloc.h>
  23 #include "internal.h"
  24
  25 /*
  26  * By default transparent hugepage support is enabled for all mappings
  27  * and khugepaged scans all mappings. Defrag is only invoked by
  28  * khugepaged hugepage allocations and by page faults inside
  29  * MADV_HUGEPAGE regions to avoid the risk of slowing down short lived
  30  * allocations.
  31  */
  32 unsigned long transparent_hugepage_flags __read_mostly =
  33 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
  34         (1<<TRANSPARENT_HUGEPAGE_FLAG)|
  35 #endif
  36 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
  37         (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
  38 #endif
  39         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
  40         (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
  41
  42 /* default scan 8*512 pte (or vmas) every 30 second */
  43 static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
  44 static unsigned int khugepaged_pages_collapsed;
  45 static unsigned int khugepaged_full_scans;
  46 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  47 /* during fragmentation poll the hugepage allocator once every minute */
  48 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  49 static struct task_struct *khugepaged_thread __read_mostly;
  50 static unsigned long huge_zero_pfn __read_mostly;
  51 static DEFINE_MUTEX(khugepaged_mutex);
  52 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  53 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  54 /*
  55  * default collapse hugepages if there is at least one pte mapped like
  56  * it would have happened if the vma was large enough during page
  57  * fault.
  58  */
  59 static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
  60
  61 static int khugepaged(void *none);
  62 static int mm_slots_hash_init(void);
  63 static int khugepaged_slab_init(void);
  64 static void khugepaged_slab_free(void);
  65
  66 #define MM_SLOTS_HASH_HEADS 1024
  67 static struct hlist_head *mm_slots_hash __read_mostly;
  68 static struct kmem_cache *mm_slot_cache __read_mostly;
  69
  70 /**
  71  * struct mm_slot - hash lookup from mm to mm_slot
  72  * @hash: hash collision list
  73  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  74  * @mm: the mm that this information is valid for
  75  */
  76 struct mm_slot {
  77         struct hlist_node hash;
  78         struct list_head mm_node;
  79         struct mm_struct *mm;
  80 };
  81
  82 /**
  83  * struct khugepaged_scan - cursor for scanning
  84  * @mm_head: the head of the mm list to scan
  85  * @mm_slot: the current mm_slot we are scanning
  86  * @address: the next address inside that to be scanned
  87  *
  88  * There is only the one khugepaged_scan instance of this cursor structure.
  89  */
  90 struct khugepaged_scan {
  91         struct list_head mm_head;
  92         struct mm_slot *mm_slot;
  93         unsigned long address;
  94 };
  95 static struct khugepaged_scan khugepaged_scan = {
  96         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
  97 };
  98
  99
 100 static int set_recommended_min_free_kbytes(void)
 101 {
 102         struct zone *zone;
 103         int nr_zones = 0;
 104         unsigned long recommended_min;
 105         extern int min_free_kbytes;
 106
 107         if (!khugepaged_enabled())
 108                 return 0;
 109
 110         for_each_populated_zone(zone)
 111                 nr_zones++;
 112
 113         /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
 114         recommended_min = pageblock_nr_pages * nr_zones * 2;
 115
 116         /*
 117          * Make sure that on average at least two pageblocks are almost free
 118          * of another type, one for a migratetype to fall back to and a
 119          * second to avoid subsequent fallbacks of other types There are 3
 120          * MIGRATE_TYPES we care about.
 121          */
 122         recommended_min += pageblock_nr_pages * nr_zones *
 123                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
 124
 125         /* don't ever allow to reserve more than 5% of the lowmem */
 126         recommended_min = min(recommended_min,
 127                               (unsigned long) nr_free_buffer_pages() / 20);
 128         recommended_min <<= (PAGE_SHIFT-10);
 129
 130         if (recommended_min > min_free_kbytes)
 131                 min_free_kbytes = recommended_min;
 132         setup_per_zone_wmarks();
 133         return 0;
 134 }
 135 late_initcall(set_recommended_min_free_kbytes);
 136
 137 static int start_khugepaged(void)
 138 {
 139         int err = 0;
 140         if (khugepaged_enabled()) {
 141                 if (!khugepaged_thread)
 142                         khugepaged_thread = kthread_run(khugepaged, NULL,
 143                                                         "khugepaged");
 144                 if (unlikely(IS_ERR(khugepaged_thread))) {
 145                         printk(KERN_ERR
 146                                "khugepaged: kthread_run(khugepaged) failed\n");
 147                         err = PTR_ERR(khugepaged_thread);
 148                         khugepaged_thread = NULL;
 149                 }
 150
 151                 if (!list_empty(&khugepaged_scan.mm_head))
 152                         wake_up_interruptible(&khugepaged_wait);
 153
 154                 set_recommended_min_free_kbytes();
 155         } else if (khugepaged_thread) {
 156                 kthread_stop(khugepaged_thread);
 157                 khugepaged_thread = NULL;
 158         }
 159
 160         return err;
 161 }
 162
 163 static int __init init_huge_zero_page(void)
 164 {
 165         struct page *hpage;
 166
 167         hpage = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
 168                         HPAGE_PMD_ORDER);
 169         if (!hpage)
 170                 return -ENOMEM;
 171
 172         huge_zero_pfn = page_to_pfn(hpage);
 173         return 0;
 174 }
 175
 176 static inline bool is_huge_zero_pfn(unsigned long pfn)
 177 {
 178         return pfn == huge_zero_pfn;
 179 }
 180
 181 static inline bool is_huge_zero_pmd(pmd_t pmd)
 182 {
 183         return is_huge_zero_pfn(pmd_pfn(pmd));
 184 }
 185
 186 #ifdef CONFIG_SYSFS
 187
 188 static ssize_t double_flag_show(struct kobject *kobj,
 189                                 struct kobj_attribute *attr, char *buf,
 190                                 enum transparent_hugepage_flag enabled,
 191                                 enum transparent_hugepage_flag req_madv)
 192 {
 193         if (test_bit(enabled, &transparent_hugepage_flags)) {
 194                 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
 195                 return sprintf(buf, "[always] madvise never\n");
 196         } else if (test_bit(req_madv, &transparent_hugepage_flags))
 197                 return sprintf(buf, "always [madvise] never\n");
 198         else
 199                 return sprintf(buf, "always madvise [never]\n");
 200 }
 201 static ssize_t double_flag_store(struct kobject *kobj,
 202                                  struct kobj_attribute *attr,
 203                                  const char *buf, size_t count,
 204                                  enum transparent_hugepage_flag enabled,
 205                                  enum transparent_hugepage_flag req_madv)
 206 {
 207         if (!memcmp("always", buf,
 208                     min(sizeof("always")-1, count))) {
 209                 set_bit(enabled, &transparent_hugepage_flags);
 210                 clear_bit(req_madv, &transparent_hugepage_flags);
 211         } else if (!memcmp("madvise", buf,
 212                            min(sizeof("madvise")-1, count))) {
 213                 clear_bit(enabled, &transparent_hugepage_flags);
 214                 set_bit(req_madv, &transparent_hugepage_flags);
 215         } else if (!memcmp("never", buf,
 216                            min(sizeof("never")-1, count))) {
 217                 clear_bit(enabled, &transparent_hugepage_flags);
 218                 clear_bit(req_madv, &transparent_hugepage_flags);
 219         } else
 220                 return -EINVAL;
 221
 222         return count;
 223 }
 224
 225 static ssize_t enabled_show(struct kobject *kobj,
 226                             struct kobj_attribute *attr, char *buf)
 227 {
 228         return double_flag_show(kobj, attr, buf,
 229                                 TRANSPARENT_HUGEPAGE_FLAG,
 230                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 231 }
 232 static ssize_t enabled_store(struct kobject *kobj,
 233                              struct kobj_attribute *attr,
 234                              const char *buf, size_t count)
 235 {
 236         ssize_t ret;
 237
 238         ret = double_flag_store(kobj, attr, buf, count,
 239                                 TRANSPARENT_HUGEPAGE_FLAG,
 240                                 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
 241
 242         if (ret > 0) {
 243                 int err;
 244
 245                 mutex_lock(&khugepaged_mutex);
 246                 err = start_khugepaged();
 247                 mutex_unlock(&khugepaged_mutex);
 248
 249                 if (err)
 250                         ret = err;
 251         }
 252
 253         return ret;
 254 }
 255 static struct kobj_attribute enabled_attr =
 256         __ATTR(enabled, 0644, enabled_show, enabled_store);
 257
 258 static ssize_t single_flag_show(struct kobject *kobj,
 259                                 struct kobj_attribute *attr, char *buf,
 260                                 enum transparent_hugepage_flag flag)
 261 {
 262         return sprintf(buf, "%d\n",
 263                        !!test_bit(flag, &transparent_hugepage_flags));
 264 }
 265
 266 static ssize_t single_flag_store(struct kobject *kobj,
 267                                  struct kobj_attribute *attr,
 268                                  const char *buf, size_t count,
 269                                  enum transparent_hugepage_flag flag)
 270 {
 271         unsigned long value;
 272         int ret;
 273
 274         ret = kstrtoul(buf, 10, &value);
 275         if (ret < 0)
 276                 return ret;
 277         if (value > 1)
 278                 return -EINVAL;
 279
 280         if (value)
 281                 set_bit(flag, &transparent_hugepage_flags);
 282         else
 283                 clear_bit(flag, &transparent_hugepage_flags);
 284
 285         return count;
 286 }
 287
 288 /*
 289  * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
 290  * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
 291  * memory just to allocate one more hugepage.
 292  */
 293 static ssize_t defrag_show(struct kobject *kobj,
 294                            struct kobj_attribute *attr, char *buf)
 295 {
 296         return double_flag_show(kobj, attr, buf,
 297                                 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 298                                 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 299 }
 300 static ssize_t defrag_store(struct kobject *kobj,
 301                             struct kobj_attribute *attr,
 302                             const char *buf, size_t count)
 303 {
 304         return double_flag_store(kobj, attr, buf, count,
 305                                  TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
 306                                  TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
 307 }
 308 static struct kobj_attribute defrag_attr =
 309         __ATTR(defrag, 0644, defrag_show, defrag_store);
 310
 311 #ifdef CONFIG_DEBUG_VM
 312 static ssize_t debug_cow_show(struct kobject *kobj,
 313                                 struct kobj_attribute *attr, char *buf)
 314 {
 315         return single_flag_show(kobj, attr, buf,
 316                                 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 317 }
 318 static ssize_t debug_cow_store(struct kobject *kobj,
 319                                struct kobj_attribute *attr,
 320                                const char *buf, size_t count)
 321 {
 322         return single_flag_store(kobj, attr, buf, count,
 323                                  TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
 324 }
 325 static struct kobj_attribute debug_cow_attr =
 326         __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
 327 #endif /* CONFIG_DEBUG_VM */
 328
 329 static struct attribute *hugepage_attr[] = {
 330         &enabled_attr.attr,
 331         &defrag_attr.attr,
 332 #ifdef CONFIG_DEBUG_VM
 333         &debug_cow_attr.attr,
 334 #endif
 335         NULL,
 336 };
 337
 338 static struct attribute_group hugepage_attr_group = {
 339         .attrs = hugepage_attr,
 340 };
 341
 342 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 343                                          struct kobj_attribute *attr,
 344                                          char *buf)
 345 {
 346         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 347 }
 348
 349 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 350                                           struct kobj_attribute *attr,
 351                                           const char *buf, size_t count)
 352 {
 353         unsigned long msecs;
 354         int err;
 355
 356         err = strict_strtoul(buf, 10, &msecs);
 357         if (err || msecs > UINT_MAX)
 358                 return -EINVAL;
 359
 360         khugepaged_scan_sleep_millisecs = msecs;
 361         wake_up_interruptible(&khugepaged_wait);
 362
 363         return count;
 364 }
 365 static struct kobj_attribute scan_sleep_millisecs_attr =
 366         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 367                scan_sleep_millisecs_store);
 368
 369 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 370                                           struct kobj_attribute *attr,
 371                                           char *buf)
 372 {
 373         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 374 }
 375
 376 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 377                                            struct kobj_attribute *attr,
 378                                            const char *buf, size_t count)
 379 {
 380         unsigned long msecs;
 381         int err;
 382
 383         err = strict_strtoul(buf, 10, &msecs);
 384         if (err || msecs > UINT_MAX)
 385                 return -EINVAL;
 386
 387         khugepaged_alloc_sleep_millisecs = msecs;
 388         wake_up_interruptible(&khugepaged_wait);
 389
 390         return count;
 391 }
 392 static struct kobj_attribute alloc_sleep_millisecs_attr =
 393         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 394                alloc_sleep_millisecs_store);
 395
 396 static ssize_t pages_to_scan_show(struct kobject *kobj,
 397                                   struct kobj_attribute *attr,
 398                                   char *buf)
 399 {
 400         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 401 }
 402 static ssize_t pages_to_scan_store(struct kobject *kobj,
 403                                    struct kobj_attribute *attr,
 404                                    const char *buf, size_t count)
 405 {
 406         int err;
 407         unsigned long pages;
 408
 409         err = strict_strtoul(buf, 10, &pages);
 410         if (err || !pages || pages > UINT_MAX)
 411                 return -EINVAL;
 412
 413         khugepaged_pages_to_scan = pages;
 414
 415         return count;
 416 }
 417 static struct kobj_attribute pages_to_scan_attr =
 418         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 419                pages_to_scan_store);
 420
 421 static ssize_t pages_collapsed_show(struct kobject *kobj,
 422                                     struct kobj_attribute *attr,
 423                                     char *buf)
 424 {
 425         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 426 }
 427 static struct kobj_attribute pages_collapsed_attr =
 428         __ATTR_RO(pages_collapsed);
 429
 430 static ssize_t full_scans_show(struct kobject *kobj,
 431                                struct kobj_attribute *attr,
 432                                char *buf)
 433 {
 434         return sprintf(buf, "%u\n", khugepaged_full_scans);
 435 }
 436 static struct kobj_attribute full_scans_attr =
 437         __ATTR_RO(full_scans);
 438
 439 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 440                                       struct kobj_attribute *attr, char *buf)
 441 {
 442         return single_flag_show(kobj, attr, buf,
 443                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 444 }
 445 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 446                                        struct kobj_attribute *attr,
 447                                        const char *buf, size_t count)
 448 {
 449         return single_flag_store(kobj, attr, buf, count,
 450                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 451 }
 452 static struct kobj_attribute khugepaged_defrag_attr =
 453         __ATTR(defrag, 0644, khugepaged_defrag_show,
 454                khugepaged_defrag_store);
 455
 456 /*
 457  * max_ptes_none controls if khugepaged should collapse hugepages over
 458  * any unmapped ptes in turn potentially increasing the memory
 459  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 460  * reduce the available free memory in the system as it
 461  * runs. Increasing max_ptes_none will instead potentially reduce the
 462  * free memory in the system during the khugepaged scan.
 463  */
 464 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 465                                              struct kobj_attribute *attr,
 466                                              char *buf)
 467 {
 468         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 469 }
 470 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 471                                               struct kobj_attribute *attr,
 472                                               const char *buf, size_t count)
 473 {
 474         int err;
 475         unsigned long max_ptes_none;
 476
 477         err = strict_strtoul(buf, 10, &max_ptes_none);
 478         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 479                 return -EINVAL;
 480
 481         khugepaged_max_ptes_none = max_ptes_none;
 482
 483         return count;
 484 }
 485 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 486         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 487                khugepaged_max_ptes_none_store);
 488
 489 static struct attribute *khugepaged_attr[] = {
 490         &khugepaged_defrag_attr.attr,
 491         &khugepaged_max_ptes_none_attr.attr,
 492         &pages_to_scan_attr.attr,
 493         &pages_collapsed_attr.attr,
 494         &full_scans_attr.attr,
 495         &scan_sleep_millisecs_attr.attr,
 496         &alloc_sleep_millisecs_attr.attr,
 497         NULL,
 498 };
 499
 500 static struct attribute_group khugepaged_attr_group = {
 501         .attrs = khugepaged_attr,
 502         .name = "khugepaged",
 503 };
 504
 505 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
 506 {
 507         int err;
 508
 509         *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
 510         if (unlikely(!*hugepage_kobj)) {
 511                 printk(KERN_ERR "hugepage: failed kobject create\n");
 512                 return -ENOMEM;
 513         }
 514
 515         err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
 516         if (err) {
 517                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 518                 goto delete_obj;
 519         }
 520
 521         err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
 522         if (err) {
 523                 printk(KERN_ERR "hugepage: failed register hugeage group\n");
 524                 goto remove_hp_group;
 525         }
 526
 527         return 0;
 528
 529 remove_hp_group:
 530         sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
 531 delete_obj:
 532         kobject_put(*hugepage_kobj);
 533         return err;
 534 }
 535
 536 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 537 {
 538         sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
 539         sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
 540         kobject_put(hugepage_kobj);
 541 }
 542 #else
 543 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
 544 {
 545         return 0;
 546 }
 547
 548 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
 549 {
 550 }
 551 #endif /* CONFIG_SYSFS */
 552
 553 static int __init hugepage_init(void)
 554 {
 555         int err;
 556         struct kobject *hugepage_kobj;
 557
 558         if (!has_transparent_hugepage()) {
 559                 transparent_hugepage_flags = 0;
 560                 return -EINVAL;
 561         }
 562
 563         err = hugepage_init_sysfs(&hugepage_kobj);
 564         if (err)
 565                 return err;
 566
 567         err = init_huge_zero_page();
 568         if (err)
 569                 goto out;
 570
 571         err = khugepaged_slab_init();
 572         if (err)
 573                 goto out;
 574
 575         err = mm_slots_hash_init();
 576         if (err) {
 577                 khugepaged_slab_free();
 578                 goto out;
 579         }
 580
 581         /*
 582          * By default disable transparent hugepages on smaller systems,
 583          * where the extra memory used could hurt more than TLB overhead
 584          * is likely to save.  The admin can still enable it through /sys.
 585          */
 586         if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
 587                 transparent_hugepage_flags = 0;
 588
 589         start_khugepaged();
 590
 591         return 0;
 592 out:
 593         if (huge_zero_pfn)
 594                 __free_page(pfn_to_page(huge_zero_pfn));
 595         hugepage_exit_sysfs(hugepage_kobj);
 596         return err;
 597 }
 598 module_init(hugepage_init)
 599
 600 static int __init setup_transparent_hugepage(char *str)
 601 {
 602         int ret = 0;
 603         if (!str)
 604                 goto out;
 605         if (!strcmp(str, "always")) {
 606                 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
 607                         &transparent_hugepage_flags);
 608                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 609                           &transparent_hugepage_flags);
 610                 ret = 1;
 611         } else if (!strcmp(str, "madvise")) {
 612                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 613                           &transparent_hugepage_flags);
 614                 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 615                         &transparent_hugepage_flags);
 616                 ret = 1;
 617         } else if (!strcmp(str, "never")) {
 618                 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
 619                           &transparent_hugepage_flags);
 620                 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
 621                           &transparent_hugepage_flags);
 622                 ret = 1;
 623         }
 624 out:
 625         if (!ret)
 626                 printk(KERN_WARNING
 627                        "transparent_hugepage= cannot parse, ignored\n");
 628         return ret;
 629 }
 630 __setup("transparent_hugepage=", setup_transparent_hugepage);
 631
 632 static inline pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
 633 {
 634         if (likely(vma->vm_flags & VM_WRITE))
 635                 pmd = pmd_mkwrite(pmd);
 636         return pmd;
 637 }
 638
 639 static inline pmd_t mk_huge_pmd(struct page *page, struct vm_area_struct *vma)
 640 {
 641         pmd_t entry;
 642         entry = mk_pmd(page, vma->vm_page_prot);
 643         entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
 644         entry = pmd_mkhuge(entry);
 645         return entry;
 646 }
 647
 648 static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
 649                                         struct vm_area_struct *vma,
 650                                         unsigned long haddr, pmd_t *pmd,
 651                                         struct page *page)
 652 {
 653         pgtable_t pgtable;
 654
 655         VM_BUG_ON(!PageCompound(page));
 656         pgtable = pte_alloc_one(mm, haddr);
 657         if (unlikely(!pgtable))
 658                 return VM_FAULT_OOM;
 659
 660         clear_huge_page(page, haddr, HPAGE_PMD_NR);
 661         __SetPageUptodate(page);
 662
 663         spin_lock(&mm->page_table_lock);
 664         if (unlikely(!pmd_none(*pmd))) {
 665                 spin_unlock(&mm->page_table_lock);
 666                 mem_cgroup_uncharge_page(page);
 667                 put_page(page);
 668                 pte_free(mm, pgtable);
 669         } else {
 670                 pmd_t entry;
 671                 entry = mk_huge_pmd(page, vma);
 672                 /*
 673                  * The spinlocking to take the lru_lock inside
 674                  * page_add_new_anon_rmap() acts as a full memory
 675                  * barrier to be sure clear_huge_page writes become
 676                  * visible after the set_pmd_at() write.
 677                  */
 678                 page_add_new_anon_rmap(page, vma, haddr);
 679                 set_pmd_at(mm, haddr, pmd, entry);
 680                 pgtable_trans_huge_deposit(mm, pgtable);
 681                 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
 682                 mm->nr_ptes++;
 683                 spin_unlock(&mm->page_table_lock);
 684         }
 685
 686         return 0;
 687 }
 688
 689 static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
 690 {
 691         return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
 692 }
 693
 694 static inline struct page *alloc_hugepage_vma(int defrag,
 695                                               struct vm_area_struct *vma,
 696                                               unsigned long haddr, int nd,
 697                                               gfp_t extra_gfp)
 698 {
 699         return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
 700                                HPAGE_PMD_ORDER, vma, haddr, nd);
 701 }
 702
 703 #ifndef CONFIG_NUMA
 704 static inline struct page *alloc_hugepage(int defrag)
 705 {
 706         return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
 707                            HPAGE_PMD_ORDER);
 708 }
 709 #endif
 710
 711 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
 712                 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd)
 713 {
 714         pmd_t entry;
 715         entry = pfn_pmd(huge_zero_pfn, vma->vm_page_prot);
 716         entry = pmd_wrprotect(entry);
 717         entry = pmd_mkhuge(entry);
 718         set_pmd_at(mm, haddr, pmd, entry);
 719         pgtable_trans_huge_deposit(mm, pgtable);
 720         mm->nr_ptes++;
 721 }
 722
 723 int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 724                                unsigned long address, pmd_t *pmd,
 725                                unsigned int flags)
 726 {
 727         struct page *page;
 728         unsigned long haddr = address & HPAGE_PMD_MASK;
 729         pte_t *pte;
 730
 731         if (haddr >= vma->vm_start && haddr + HPAGE_PMD_SIZE <= vma->vm_end) {
 732                 if (unlikely(anon_vma_prepare(vma)))
 733                         return VM_FAULT_OOM;
 734                 if (unlikely(khugepaged_enter(vma)))
 735                         return VM_FAULT_OOM;
 736                 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
 737                                           vma, haddr, numa_node_id(), 0);
 738                 if (unlikely(!page)) {
 739                         count_vm_event(THP_FAULT_FALLBACK);
 740                         goto out;
 741                 }
 742                 count_vm_event(THP_FAULT_ALLOC);
 743                 if (unlikely(mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))) {
 744                         put_page(page);
 745                         goto out;
 746                 }
 747                 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd,
 748                                                           page))) {
 749                         mem_cgroup_uncharge_page(page);
 750                         put_page(page);
 751                         goto out;
 752                 }
 753
 754                 return 0;
 755         }
 756 out:
 757         /*
 758          * Use __pte_alloc instead of pte_alloc_map, because we can't
 759          * run pte_offset_map on the pmd, if an huge pmd could
 760          * materialize from under us from a different thread.
 761          */
 762         if (unlikely(__pte_alloc(mm, vma, pmd, address)))
 763                 return VM_FAULT_OOM;
 764         /* if an huge pmd materialized from under us just retry later */
 765         if (unlikely(pmd_trans_huge(*pmd)))
 766                 return 0;
 767         /*
 768          * A regular pmd is established and it can't morph into a huge pmd
 769          * from under us anymore at this point because we hold the mmap_sem
 770          * read mode and khugepaged takes it in write mode. So now it's
 771          * safe to run pte_offset_map().
 772          */
 773         pte = pte_offset_map(pmd, address);
 774         return handle_pte_fault(mm, vma, address, pte, pmd, flags);
 775 }
 776
 777 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
 778                   pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
 779                   struct vm_area_struct *vma)
 780 {
 781         struct page *src_page;
 782         pmd_t pmd;
 783         pgtable_t pgtable;
 784         int ret;
 785
 786         ret = -ENOMEM;
 787         pgtable = pte_alloc_one(dst_mm, addr);
 788         if (unlikely(!pgtable))
 789                 goto out;
 790
 791         spin_lock(&dst_mm->page_table_lock);
 792         spin_lock_nested(&src_mm->page_table_lock, SINGLE_DEPTH_NESTING);
 793
 794         ret = -EAGAIN;
 795         pmd = *src_pmd;
 796         if (unlikely(!pmd_trans_huge(pmd))) {
 797                 pte_free(dst_mm, pgtable);
 798                 goto out_unlock;
 799         }
 800         /*
 801          * mm->page_table_lock is enough to be sure that huge zero pmd is not
 802          * under splitting since we don't split the page itself, only pmd to
 803          * a page table.
 804          */
 805         if (is_huge_zero_pmd(pmd)) {
 806                 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd);
 807                 ret = 0;
 808                 goto out_unlock;
 809         }
 810         if (unlikely(pmd_trans_splitting(pmd))) {
 811                 /* split huge page running from under us */
 812                 spin_unlock(&src_mm->page_table_lock);
 813                 spin_unlock(&dst_mm->page_table_lock);
 814                 pte_free(dst_mm, pgtable);
 815
 816                 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
 817                 goto out;
 818         }
 819         src_page = pmd_page(pmd);
 820         VM_BUG_ON(!PageHead(src_page));
 821         get_page(src_page);
 822         page_dup_rmap(src_page);
 823         add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
 824
 825         pmdp_set_wrprotect(src_mm, addr, src_pmd);
 826         pmd = pmd_mkold(pmd_wrprotect(pmd));
 827         set_pmd_at(dst_mm, addr, dst_pmd, pmd);
 828         pgtable_trans_huge_deposit(dst_mm, pgtable);
 829         dst_mm->nr_ptes++;
 830
 831         ret = 0;
 832 out_unlock:
 833         spin_unlock(&src_mm->page_table_lock);
 834         spin_unlock(&dst_mm->page_table_lock);
 835 out:
 836         return ret;
 837 }
 838
 839 void huge_pmd_set_accessed(struct mm_struct *mm,
 840                            struct vm_area_struct *vma,
 841                            unsigned long address,
 842                            pmd_t *pmd, pmd_t orig_pmd,
 843                            int dirty)
 844 {
 845         pmd_t entry;
 846         unsigned long haddr;
 847
 848         spin_lock(&mm->page_table_lock);
 849         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 850                 goto unlock;
 851
 852         entry = pmd_mkyoung(orig_pmd);
 853         haddr = address & HPAGE_PMD_MASK;
 854         if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
 855                 update_mmu_cache_pmd(vma, address, pmd);
 856
 857 unlock:
 858         spin_unlock(&mm->page_table_lock);
 859 }
 860
 861 static int do_huge_pmd_wp_zero_page_fallback(struct mm_struct *mm,
 862                 struct vm_area_struct *vma, unsigned long address,
 863                 pmd_t *pmd, unsigned long haddr)
 864 {
 865         pgtable_t pgtable;
 866         pmd_t _pmd;
 867         struct page *page;
 868         int i, ret = 0;
 869         unsigned long mmun_start;       /* For mmu_notifiers */
 870         unsigned long mmun_end;         /* For mmu_notifiers */
 871
 872         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
 873         if (!page) {
 874                 ret |= VM_FAULT_OOM;
 875                 goto out;
 876         }
 877
 878         if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
 879                 put_page(page);
 880                 ret |= VM_FAULT_OOM;
 881                 goto out;
 882         }
 883
 884         clear_user_highpage(page, address);
 885         __SetPageUptodate(page);
 886
 887         mmun_start = haddr;
 888         mmun_end   = haddr + HPAGE_PMD_SIZE;
 889         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 890
 891         spin_lock(&mm->page_table_lock);
 892         pmdp_clear_flush(vma, haddr, pmd);
 893         /* leave pmd empty until pte is filled */
 894
 895         pgtable = pgtable_trans_huge_withdraw(mm);
 896         pmd_populate(mm, &_pmd, pgtable);
 897
 898         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 899                 pte_t *pte, entry;
 900                 if (haddr == (address & PAGE_MASK)) {
 901                         entry = mk_pte(page, vma->vm_page_prot);
 902                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 903                         page_add_new_anon_rmap(page, vma, haddr);
 904                 } else {
 905                         entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
 906                         entry = pte_mkspecial(entry);
 907                 }
 908                 pte = pte_offset_map(&_pmd, haddr);
 909                 VM_BUG_ON(!pte_none(*pte));
 910                 set_pte_at(mm, haddr, pte, entry);
 911                 pte_unmap(pte);
 912         }
 913         smp_wmb(); /* make pte visible before pmd */
 914         pmd_populate(mm, pmd, pgtable);
 915         spin_unlock(&mm->page_table_lock);
 916         inc_mm_counter(mm, MM_ANONPAGES);
 917
 918         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
 919
 920         ret |= VM_FAULT_WRITE;
 921 out:
 922         return ret;
 923 }
 924
 925 static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
 926                                         struct vm_area_struct *vma,
 927                                         unsigned long address,
 928                                         pmd_t *pmd, pmd_t orig_pmd,
 929                                         struct page *page,
 930                                         unsigned long haddr)
 931 {
 932         pgtable_t pgtable;
 933         pmd_t _pmd;
 934         int ret = 0, i;
 935         struct page **pages;
 936         unsigned long mmun_start;       /* For mmu_notifiers */
 937         unsigned long mmun_end;         /* For mmu_notifiers */
 938
 939         pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
 940                         GFP_KERNEL);
 941         if (unlikely(!pages)) {
 942                 ret |= VM_FAULT_OOM;
 943                 goto out;
 944         }
 945
 946         for (i = 0; i < HPAGE_PMD_NR; i++) {
 947                 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
 948                                                __GFP_OTHER_NODE,
 949                                                vma, address, page_to_nid(page));
 950                 if (unlikely(!pages[i] ||
 951                              mem_cgroup_newpage_charge(pages[i], mm,
 952                                                        GFP_KERNEL))) {
 953                         if (pages[i])
 954                                 put_page(pages[i]);
 955                         mem_cgroup_uncharge_start();
 956                         while (--i >= 0) {
 957                                 mem_cgroup_uncharge_page(pages[i]);
 958                                 put_page(pages[i]);
 959                         }
 960                         mem_cgroup_uncharge_end();
 961                         kfree(pages);
 962                         ret |= VM_FAULT_OOM;
 963                         goto out;
 964                 }
 965         }
 966
 967         for (i = 0; i < HPAGE_PMD_NR; i++) {
 968                 copy_user_highpage(pages[i], page + i,
 969                                    haddr + PAGE_SIZE * i, vma);
 970                 __SetPageUptodate(pages[i]);
 971                 cond_resched();
 972         }
 973
 974         mmun_start = haddr;
 975         mmun_end   = haddr + HPAGE_PMD_SIZE;
 976         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
 977
 978         spin_lock(&mm->page_table_lock);
 979         if (unlikely(!pmd_same(*pmd, orig_pmd)))
 980                 goto out_free_pages;
 981         VM_BUG_ON(!PageHead(page));
 982
 983         pmdp_clear_flush(vma, haddr, pmd);
 984         /* leave pmd empty until pte is filled */
 985
 986         pgtable = pgtable_trans_huge_withdraw(mm);
 987         pmd_populate(mm, &_pmd, pgtable);
 988
 989         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
 990                 pte_t *pte, entry;
 991                 entry = mk_pte(pages[i], vma->vm_page_prot);
 992                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
 993                 page_add_new_anon_rmap(pages[i], vma, haddr);
 994                 pte = pte_offset_map(&_pmd, haddr);
 995                 VM_BUG_ON(!pte_none(*pte));
 996                 set_pte_at(mm, haddr, pte, entry);
 997                 pte_unmap(pte);
 998         }
 999         kfree(pages);
1000
1001         smp_wmb(); /* make pte visible before pmd */
1002         pmd_populate(mm, pmd, pgtable);
1003         page_remove_rmap(page);
1004         spin_unlock(&mm->page_table_lock);
1005
1006         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1007
1008         ret |= VM_FAULT_WRITE;
1009         put_page(page);
1010
1011 out:
1012         return ret;
1013
1014 out_free_pages:
1015         spin_unlock(&mm->page_table_lock);
1016         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1017         mem_cgroup_uncharge_start();
1018         for (i = 0; i < HPAGE_PMD_NR; i++) {
1019                 mem_cgroup_uncharge_page(pages[i]);
1020                 put_page(pages[i]);
1021         }
1022         mem_cgroup_uncharge_end();
1023         kfree(pages);
1024         goto out;
1025 }
1026
1027 int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1028                         unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1029 {
1030         int ret = 0;
1031         struct page *page = NULL, *new_page;
1032         unsigned long haddr;
1033         unsigned long mmun_start;       /* For mmu_notifiers */
1034         unsigned long mmun_end;         /* For mmu_notifiers */
1035
1036         VM_BUG_ON(!vma->anon_vma);
1037         haddr = address & HPAGE_PMD_MASK;
1038         if (is_huge_zero_pmd(orig_pmd))
1039                 goto alloc;
1040         spin_lock(&mm->page_table_lock);
1041         if (unlikely(!pmd_same(*pmd, orig_pmd)))
1042                 goto out_unlock;
1043
1044         page = pmd_page(orig_pmd);
1045         VM_BUG_ON(!PageCompound(page) || !PageHead(page));
1046         if (page_mapcount(page) == 1) {
1047                 pmd_t entry;
1048                 entry = pmd_mkyoung(orig_pmd);
1049                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1050                 if (pmdp_set_access_flags(vma, haddr, pmd, entry,  1))
1051                         update_mmu_cache_pmd(vma, address, pmd);
1052                 ret |= VM_FAULT_WRITE;
1053                 goto out_unlock;
1054         }
1055         get_page(page);
1056         spin_unlock(&mm->page_table_lock);
1057 alloc:
1058         if (transparent_hugepage_enabled(vma) &&
1059             !transparent_hugepage_debug_cow())
1060                 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1061                                               vma, haddr, numa_node_id(), 0);
1062         else
1063                 new_page = NULL;
1064
1065         if (unlikely(!new_page)) {
1066                 count_vm_event(THP_FAULT_FALLBACK);
1067                 if (is_huge_zero_pmd(orig_pmd)) {
1068                         ret = do_huge_pmd_wp_zero_page_fallback(mm, vma,
1069                                         address, pmd, haddr);
1070                 } else {
1071                         ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1072                                         pmd, orig_pmd, page, haddr);
1073                         if (ret & VM_FAULT_OOM)
1074                                 split_huge_page(page);
1075                         put_page(page);
1076                 }
1077                 goto out;
1078         }
1079         count_vm_event(THP_FAULT_ALLOC);
1080
1081         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))) {
1082                 put_page(new_page);
1083                 if (page) {
1084                         split_huge_page(page);
1085                         put_page(page);
1086                 }
1087                 ret |= VM_FAULT_OOM;
1088                 goto out;
1089         }
1090
1091         if (is_huge_zero_pmd(orig_pmd))
1092                 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1093         else
1094                 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1095         __SetPageUptodate(new_page);
1096
1097         mmun_start = haddr;
1098         mmun_end   = haddr + HPAGE_PMD_SIZE;
1099         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1100
1101         spin_lock(&mm->page_table_lock);
1102         if (page)
1103                 put_page(page);
1104         if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1105                 spin_unlock(&mm->page_table_lock);
1106                 mem_cgroup_uncharge_page(new_page);
1107                 put_page(new_page);
1108                 goto out_mn;
1109         } else {
1110                 pmd_t entry;
1111                 entry = mk_huge_pmd(new_page, vma);
1112                 pmdp_clear_flush(vma, haddr, pmd);
1113                 page_add_new_anon_rmap(new_page, vma, haddr);
1114                 set_pmd_at(mm, haddr, pmd, entry);
1115                 update_mmu_cache_pmd(vma, address, pmd);
1116                 if (is_huge_zero_pmd(orig_pmd))
1117                         add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1118                 else {
1119                         VM_BUG_ON(!PageHead(page));
1120                         page_remove_rmap(page);
1121                         put_page(page);
1122                 }
1123                 ret |= VM_FAULT_WRITE;
1124         }
1125         spin_unlock(&mm->page_table_lock);
1126 out_mn:
1127         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1128 out:
1129         return ret;
1130 out_unlock:
1131         spin_unlock(&mm->page_table_lock);
1132         return ret;
1133 }
1134
1135 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1136                                    unsigned long addr,
1137                                    pmd_t *pmd,
1138                                    unsigned int flags)
1139 {
1140         struct mm_struct *mm = vma->vm_mm;
1141         struct page *page = NULL;
1142
1143         assert_spin_locked(&mm->page_table_lock);
1144
1145         if (flags & FOLL_WRITE && !pmd_write(*pmd))
1146                 goto out;
1147
1148         page = pmd_page(*pmd);
1149         VM_BUG_ON(!PageHead(page));
1150         if (flags & FOLL_TOUCH) {
1151                 pmd_t _pmd;
1152                 /*
1153                  * We should set the dirty bit only for FOLL_WRITE but
1154                  * for now the dirty bit in the pmd is meaningless.
1155                  * And if the dirty bit will become meaningful and
1156                  * we'll only set it with FOLL_WRITE, an atomic
1157                  * set_bit will be required on the pmd to set the
1158                  * young bit, instead of the current set_pmd_at.
1159                  */
1160                 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1161                 set_pmd_at(mm, addr & HPAGE_PMD_MASK, pmd, _pmd);
1162         }
1163         if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1164                 if (page->mapping && trylock_page(page)) {
1165                         lru_add_drain();
1166                         if (page->mapping)
1167                                 mlock_vma_page(page);
1168                         unlock_page(page);
1169                 }
1170         }
1171         page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1172         VM_BUG_ON(!PageCompound(page));
1173         if (flags & FOLL_GET)
1174                 get_page_foll(page);
1175
1176 out:
1177         return page;
1178 }
1179
1180 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1181                  pmd_t *pmd, unsigned long addr)
1182 {
1183         int ret = 0;
1184
1185         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1186                 struct page *page;
1187                 pgtable_t pgtable;
1188                 pmd_t orig_pmd;
1189                 pgtable = pgtable_trans_huge_withdraw(tlb->mm);
1190                 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1191                 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1192                 if (is_huge_zero_pmd(orig_pmd)) {
1193                         tlb->mm->nr_ptes--;
1194                         spin_unlock(&tlb->mm->page_table_lock);
1195                 } else {
1196                         page = pmd_page(orig_pmd);
1197                         page_remove_rmap(page);
1198                         VM_BUG_ON(page_mapcount(page) < 0);
1199                         add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1200                         VM_BUG_ON(!PageHead(page));
1201                         tlb->mm->nr_ptes--;
1202                         spin_unlock(&tlb->mm->page_table_lock);
1203                         tlb_remove_page(tlb, page);
1204                 }
1205                 pte_free(tlb->mm, pgtable);
1206                 ret = 1;
1207         }
1208         return ret;
1209 }
1210
1211 int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1212                 unsigned long addr, unsigned long end,
1213                 unsigned char *vec)
1214 {
1215         int ret = 0;
1216
1217         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1218                 /*
1219                  * All logical pages in the range are present
1220                  * if backed by a huge page.
1221                  */
1222                 spin_unlock(&vma->vm_mm->page_table_lock);
1223                 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1224                 ret = 1;
1225         }
1226
1227         return ret;
1228 }
1229
1230 int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1231                   unsigned long old_addr,
1232                   unsigned long new_addr, unsigned long old_end,
1233                   pmd_t *old_pmd, pmd_t *new_pmd)
1234 {
1235         int ret = 0;
1236         pmd_t pmd;
1237
1238         struct mm_struct *mm = vma->vm_mm;
1239
1240         if ((old_addr & ~HPAGE_PMD_MASK) ||
1241             (new_addr & ~HPAGE_PMD_MASK) ||
1242             old_end - old_addr < HPAGE_PMD_SIZE ||
1243             (new_vma->vm_flags & VM_NOHUGEPAGE))
1244                 goto out;
1245
1246         /*
1247          * The destination pmd shouldn't be established, free_pgtables()
1248          * should have release it.
1249          */
1250         if (WARN_ON(!pmd_none(*new_pmd))) {
1251                 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1252                 goto out;
1253         }
1254
1255         ret = __pmd_trans_huge_lock(old_pmd, vma);
1256         if (ret == 1) {
1257                 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1258                 VM_BUG_ON(!pmd_none(*new_pmd));
1259                 set_pmd_at(mm, new_addr, new_pmd, pmd);
1260                 spin_unlock(&mm->page_table_lock);
1261         }
1262 out:
1263         return ret;
1264 }
1265
1266 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1267                 unsigned long addr, pgprot_t newprot)
1268 {
1269         struct mm_struct *mm = vma->vm_mm;
1270         int ret = 0;
1271
1272         if (__pmd_trans_huge_lock(pmd, vma) == 1) {
1273                 pmd_t entry;
1274                 entry = pmdp_get_and_clear(mm, addr, pmd);
1275                 entry = pmd_modify(entry, newprot);
1276                 BUG_ON(pmd_write(entry));
1277                 set_pmd_at(mm, addr, pmd, entry);
1278                 spin_unlock(&vma->vm_mm->page_table_lock);
1279                 ret = 1;
1280         }
1281
1282         return ret;
1283 }
1284
1285 /*
1286  * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1287  * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1288  *
1289  * Note that if it returns 1, this routine returns without unlocking page
1290  * table locks. So callers must unlock them.
1291  */
1292 int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1293 {
1294         spin_lock(&vma->vm_mm->page_table_lock);
1295         if (likely(pmd_trans_huge(*pmd))) {
1296                 if (unlikely(pmd_trans_splitting(*pmd))) {
1297                         spin_unlock(&vma->vm_mm->page_table_lock);
1298                         wait_split_huge_page(vma->anon_vma, pmd);
1299                         return -1;
1300                 } else {
1301                         /* Thp mapped by 'pmd' is stable, so we can
1302                          * handle it as it is. */
1303                         return 1;
1304                 }
1305         }
1306         spin_unlock(&vma->vm_mm->page_table_lock);
1307         return 0;
1308 }
1309
1310 pmd_t *page_check_address_pmd(struct page *page,
1311                               struct mm_struct *mm,
1312                               unsigned long address,
1313                               enum page_check_address_pmd_flag flag)
1314 {
1315         pmd_t *pmd, *ret = NULL;
1316
1317         if (address & ~HPAGE_PMD_MASK)
1318                 goto out;
1319
1320         pmd = mm_find_pmd(mm, address);
1321         if (!pmd)
1322                 goto out;
1323         if (pmd_none(*pmd))
1324                 goto out;
1325         if (pmd_page(*pmd) != page)
1326                 goto out;
1327         /*
1328          * split_vma() may create temporary aliased mappings. There is
1329          * no risk as long as all huge pmd are found and have their
1330          * splitting bit set before __split_huge_page_refcount
1331          * runs. Finding the same huge pmd more than once during the
1332          * same rmap walk is not a problem.
1333          */
1334         if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1335             pmd_trans_splitting(*pmd))
1336                 goto out;
1337         if (pmd_trans_huge(*pmd)) {
1338                 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1339                           !pmd_trans_splitting(*pmd));
1340                 ret = pmd;
1341         }
1342 out:
1343         return ret;
1344 }
1345
1346 static int __split_huge_page_splitting(struct page *page,
1347                                        struct vm_area_struct *vma,
1348                                        unsigned long address)
1349 {
1350         struct mm_struct *mm = vma->vm_mm;
1351         pmd_t *pmd;
1352         int ret = 0;
1353         /* For mmu_notifiers */
1354         const unsigned long mmun_start = address;
1355         const unsigned long mmun_end   = address + HPAGE_PMD_SIZE;
1356
1357         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1358         spin_lock(&mm->page_table_lock);
1359         pmd = page_check_address_pmd(page, mm, address,
1360                                      PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG);
1361         if (pmd) {
1362                 /*
1363                  * We can't temporarily set the pmd to null in order
1364                  * to split it, the pmd must remain marked huge at all
1365                  * times or the VM won't take the pmd_trans_huge paths
1366                  * and it won't wait on the anon_vma->root->mutex to
1367                  * serialize against split_huge_page*.
1368                  */
1369                 pmdp_splitting_flush(vma, address, pmd);
1370                 ret = 1;
1371         }
1372         spin_unlock(&mm->page_table_lock);
1373         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1374
1375         return ret;
1376 }
1377
1378 static void __split_huge_page_refcount(struct page *page)
1379 {
1380         int i;
1381         struct zone *zone = page_zone(page);
1382         struct lruvec *lruvec;
1383         int tail_count = 0;
1384
1385         /* prevent PageLRU to go away from under us, and freeze lru stats */
1386         spin_lock_irq(&zone->lru_lock);
1387         lruvec = mem_cgroup_page_lruvec(page, zone);
1388
1389         compound_lock(page);
1390         /* complete memcg works before add pages to LRU */
1391         mem_cgroup_split_huge_fixup(page);
1392
1393         for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1394                 struct page *page_tail = page + i;
1395
1396                 /* tail_page->_mapcount cannot change */
1397                 BUG_ON(page_mapcount(page_tail) < 0);
1398                 tail_count += page_mapcount(page_tail);
1399                 /* check for overflow */
1400                 BUG_ON(tail_count < 0);
1401                 BUG_ON(atomic_read(&page_tail->_count) != 0);
1402                 /*
1403                  * tail_page->_count is zero and not changing from
1404                  * under us. But get_page_unless_zero() may be running
1405                  * from under us on the tail_page. If we used
1406                  * atomic_set() below instead of atomic_add(), we
1407                  * would then run atomic_set() concurrently with
1408                  * get_page_unless_zero(), and atomic_set() is
1409                  * implemented in C not using locked ops. spin_unlock
1410                  * on x86 sometime uses locked ops because of PPro
1411                  * errata 66, 92, so unless somebody can guarantee
1412                  * atomic_set() here would be safe on all archs (and
1413                  * not only on x86), it's safer to use atomic_add().
1414                  */
1415                 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1416                            &page_tail->_count);
1417
1418                 /* after clearing PageTail the gup refcount can be released */
1419                 smp_mb();
1420
1421                 /*
1422                  * retain hwpoison flag of the poisoned tail page:
1423                  *   fix for the unsuitable process killed on Guest Machine(KVM)
1424                  *   by the memory-failure.
1425                  */
1426                 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1427                 page_tail->flags |= (page->flags &
1428                                      ((1L << PG_referenced) |
1429                                       (1L << PG_swapbacked) |
1430                                       (1L << PG_mlocked) |
1431                                       (1L << PG_uptodate)));
1432                 page_tail->flags |= (1L << PG_dirty);
1433
1434                 /* clear PageTail before overwriting first_page */
1435                 smp_wmb();
1436
1437                 /*
1438                  * __split_huge_page_splitting() already set the
1439                  * splitting bit in all pmd that could map this
1440                  * hugepage, that will ensure no CPU can alter the
1441                  * mapcount on the head page. The mapcount is only
1442                  * accounted in the head page and it has to be
1443                  * transferred to all tail pages in the below code. So
1444                  * for this code to be safe, the split the mapcount
1445                  * can't change. But that doesn't mean userland can't
1446                  * keep changing and reading the page contents while
1447                  * we transfer the mapcount, so the pmd splitting
1448                  * status is achieved setting a reserved bit in the
1449                  * pmd, not by clearing the present bit.
1450                 */
1451                 page_tail->_mapcount = page->_mapcount;
1452
1453                 BUG_ON(page_tail->mapping);
1454                 page_tail->mapping = page->mapping;
1455
1456                 page_tail->index = page->index + i;
1457
1458                 BUG_ON(!PageAnon(page_tail));
1459                 BUG_ON(!PageUptodate(page_tail));
1460                 BUG_ON(!PageDirty(page_tail));
1461                 BUG_ON(!PageSwapBacked(page_tail));
1462
1463                 lru_add_page_tail(page, page_tail, lruvec);
1464         }
1465         atomic_sub(tail_count, &page->_count);
1466         BUG_ON(atomic_read(&page->_count) <= 0);
1467
1468         __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1469         __mod_zone_page_state(zone, NR_ANON_PAGES, HPAGE_PMD_NR);
1470
1471         ClearPageCompound(page);
1472         compound_unlock(page);
1473         spin_unlock_irq(&zone->lru_lock);
1474
1475         for (i = 1; i < HPAGE_PMD_NR; i++) {
1476                 struct page *page_tail = page + i;
1477                 BUG_ON(page_count(page_tail) <= 0);
1478                 /*
1479                  * Tail pages may be freed if there wasn't any mapping
1480                  * like if add_to_swap() is running on a lru page that
1481                  * had its mapping zapped. And freeing these pages
1482                  * requires taking the lru_lock so we do the put_page
1483                  * of the tail pages after the split is complete.
1484                  */
1485                 put_page(page_tail);
1486         }
1487
1488         /*
1489          * Only the head page (now become a regular page) is required
1490          * to be pinned by the caller.
1491          */
1492         BUG_ON(page_count(page) <= 0);
1493 }
1494
1495 static int __split_huge_page_map(struct page *page,
1496                                  struct vm_area_struct *vma,
1497                                  unsigned long address)
1498 {
1499         struct mm_struct *mm = vma->vm_mm;
1500         pmd_t *pmd, _pmd;
1501         int ret = 0, i;
1502         pgtable_t pgtable;
1503         unsigned long haddr;
1504
1505         spin_lock(&mm->page_table_lock);
1506         pmd = page_check_address_pmd(page, mm, address,
1507                                      PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG);
1508         if (pmd) {
1509                 pgtable = pgtable_trans_huge_withdraw(mm);
1510                 pmd_populate(mm, &_pmd, pgtable);
1511
1512                 haddr = address;
1513                 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1514                         pte_t *pte, entry;
1515                         BUG_ON(PageCompound(page+i));
1516                         entry = mk_pte(page + i, vma->vm_page_prot);
1517                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1518                         if (!pmd_write(*pmd))
1519                                 entry = pte_wrprotect(entry);
1520                         else
1521                                 BUG_ON(page_mapcount(page) != 1);
1522                         if (!pmd_young(*pmd))
1523                                 entry = pte_mkold(entry);
1524                         pte = pte_offset_map(&_pmd, haddr);
1525                         BUG_ON(!pte_none(*pte));
1526                         set_pte_at(mm, haddr, pte, entry);
1527                         pte_unmap(pte);
1528                 }
1529
1530                 smp_wmb(); /* make pte visible before pmd */
1531                 /*
1532                  * Up to this point the pmd is present and huge and
1533                  * userland has the whole access to the hugepage
1534                  * during the split (which happens in place). If we
1535                  * overwrite the pmd with the not-huge version
1536                  * pointing to the pte here (which of course we could
1537                  * if all CPUs were bug free), userland could trigger
1538                  * a small page size TLB miss on the small sized TLB
1539                  * while the hugepage TLB entry is still established
1540                  * in the huge TLB. Some CPU doesn't like that. See
1541                  * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1542                  * Erratum 383 on page 93. Intel should be safe but is
1543                  * also warns that it's only safe if the permission
1544                  * and cache attributes of the two entries loaded in
1545                  * the two TLB is identical (which should be the case
1546                  * here). But it is generally safer to never allow
1547                  * small and huge TLB entries for the same virtual
1548                  * address to be loaded simultaneously. So instead of
1549                  * doing "pmd_populate(); flush_tlb_range();" we first
1550                  * mark the current pmd notpresent (atomically because
1551                  * here the pmd_trans_huge and pmd_trans_splitting
1552                  * must remain set at all times on the pmd until the
1553                  * split is complete for this pmd), then we flush the
1554                  * SMP TLB and finally we write the non-huge version
1555                  * of the pmd entry with pmd_populate.
1556                  */
1557                 pmdp_invalidate(vma, address, pmd);
1558                 pmd_populate(mm, pmd, pgtable);
1559                 ret = 1;
1560         }
1561         spin_unlock(&mm->page_table_lock);
1562
1563         return ret;
1564 }
1565
1566 /* must be called with anon_vma->root->mutex hold */
1567 static void __split_huge_page(struct page *page,
1568                               struct anon_vma *anon_vma)
1569 {
1570         int mapcount, mapcount2;
1571         pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1572         struct anon_vma_chain *avc;
1573
1574         BUG_ON(!PageHead(page));
1575         BUG_ON(PageTail(page));
1576
1577         mapcount = 0;
1578         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1579                 struct vm_area_struct *vma = avc->vma;
1580                 unsigned long addr = vma_address(page, vma);
1581                 BUG_ON(is_vma_temporary_stack(vma));
1582                 mapcount += __split_huge_page_splitting(page, vma, addr);
1583         }
1584         /*
1585          * It is critical that new vmas are added to the tail of the
1586          * anon_vma list. This guarantes that if copy_huge_pmd() runs
1587          * and establishes a child pmd before
1588          * __split_huge_page_splitting() freezes the parent pmd (so if
1589          * we fail to prevent copy_huge_pmd() from running until the
1590          * whole __split_huge_page() is complete), we will still see
1591          * the newly established pmd of the child later during the
1592          * walk, to be able to set it as pmd_trans_splitting too.
1593          */
1594         if (mapcount != page_mapcount(page))
1595                 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1596                        mapcount, page_mapcount(page));
1597         BUG_ON(mapcount != page_mapcount(page));
1598
1599         __split_huge_page_refcount(page);
1600
1601         mapcount2 = 0;
1602         anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1603                 struct vm_area_struct *vma = avc->vma;
1604                 unsigned long addr = vma_address(page, vma);
1605                 BUG_ON(is_vma_temporary_stack(vma));
1606                 mapcount2 += __split_huge_page_map(page, vma, addr);
1607         }
1608         if (mapcount != mapcount2)
1609                 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1610                        mapcount, mapcount2, page_mapcount(page));
1611         BUG_ON(mapcount != mapcount2);
1612 }
1613
1614 int split_huge_page(struct page *page)
1615 {
1616         struct anon_vma *anon_vma;
1617         int ret = 1;
1618
1619         BUG_ON(is_huge_zero_pfn(page_to_pfn(page)));
1620         BUG_ON(!PageAnon(page));
1621         anon_vma = page_lock_anon_vma(page);
1622         if (!anon_vma)
1623                 goto out;
1624         ret = 0;
1625         if (!PageCompound(page))
1626                 goto out_unlock;
1627
1628         BUG_ON(!PageSwapBacked(page));
1629         __split_huge_page(page, anon_vma);
1630         count_vm_event(THP_SPLIT);
1631
1632         BUG_ON(PageCompound(page));
1633 out_unlock:
1634         page_unlock_anon_vma(anon_vma);
1635 out:
1636         return ret;
1637 }
1638
1639 #define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
1640
1641 int hugepage_madvise(struct vm_area_struct *vma,
1642                      unsigned long *vm_flags, int advice)
1643 {
1644         struct mm_struct *mm = vma->vm_mm;
1645
1646         switch (advice) {
1647         case MADV_HUGEPAGE:
1648                 /*
1649                  * Be somewhat over-protective like KSM for now!
1650                  */
1651                 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1652                         return -EINVAL;
1653                 if (mm->def_flags & VM_NOHUGEPAGE)
1654                         return -EINVAL;
1655                 *vm_flags &= ~VM_NOHUGEPAGE;
1656                 *vm_flags |= VM_HUGEPAGE;
1657                 /*
1658                  * If the vma become good for khugepaged to scan,
1659                  * register it here without waiting a page fault that
1660                  * may not happen any time soon.
1661                  */
1662                 if (unlikely(khugepaged_enter_vma_merge(vma)))
1663                         return -ENOMEM;
1664                 break;
1665         case MADV_NOHUGEPAGE:
1666                 /*
1667                  * Be somewhat over-protective like KSM for now!
1668                  */
1669                 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1670                         return -EINVAL;
1671                 *vm_flags &= ~VM_HUGEPAGE;
1672                 *vm_flags |= VM_NOHUGEPAGE;
1673                 /*
1674                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1675                  * this vma even if we leave the mm registered in khugepaged if
1676                  * it got registered before VM_NOHUGEPAGE was set.
1677                  */
1678                 break;
1679         }
1680
1681         return 0;
1682 }
1683
1684 static int __init khugepaged_slab_init(void)
1685 {
1686         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1687                                           sizeof(struct mm_slot),
1688                                           __alignof__(struct mm_slot), 0, NULL);
1689         if (!mm_slot_cache)
1690                 return -ENOMEM;
1691
1692         return 0;
1693 }
1694
1695 static void __init khugepaged_slab_free(void)
1696 {
1697         kmem_cache_destroy(mm_slot_cache);
1698         mm_slot_cache = NULL;
1699 }
1700
1701 static inline struct mm_slot *alloc_mm_slot(void)
1702 {
1703         if (!mm_slot_cache)     /* initialization failed */
1704                 return NULL;
1705         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1706 }
1707
1708 static inline void free_mm_slot(struct mm_slot *mm_slot)
1709 {
1710         kmem_cache_free(mm_slot_cache, mm_slot);
1711 }
1712
1713 static int __init mm_slots_hash_init(void)
1714 {
1715         mm_slots_hash = kzalloc(MM_SLOTS_HASH_HEADS * sizeof(struct hlist_head),
1716                                 GFP_KERNEL);
1717         if (!mm_slots_hash)
1718                 return -ENOMEM;
1719         return 0;
1720 }
1721
1722 #if 0
1723 static void __init mm_slots_hash_free(void)
1724 {
1725         kfree(mm_slots_hash);
1726         mm_slots_hash = NULL;
1727 }
1728 #endif
1729
1730 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1731 {
1732         struct mm_slot *mm_slot;
1733         struct hlist_head *bucket;
1734         struct hlist_node *node;
1735
1736         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1737                                 % MM_SLOTS_HASH_HEADS];
1738         hlist_for_each_entry(mm_slot, node, bucket, hash) {
1739                 if (mm == mm_slot->mm)
1740                         return mm_slot;
1741         }
1742         return NULL;
1743 }
1744
1745 static void insert_to_mm_slots_hash(struct mm_struct *mm,
1746                                     struct mm_slot *mm_slot)
1747 {
1748         struct hlist_head *bucket;
1749
1750         bucket = &mm_slots_hash[((unsigned long)mm / sizeof(struct mm_struct))
1751                                 % MM_SLOTS_HASH_HEADS];
1752         mm_slot->mm = mm;
1753         hlist_add_head(&mm_slot->hash, bucket);
1754 }
1755
1756 static inline int khugepaged_test_exit(struct mm_struct *mm)
1757 {
1758         return atomic_read(&mm->mm_users) == 0;
1759 }
1760
1761 int __khugepaged_enter(struct mm_struct *mm)
1762 {
1763         struct mm_slot *mm_slot;
1764         int wakeup;
1765
1766         mm_slot = alloc_mm_slot();
1767         if (!mm_slot)
1768                 return -ENOMEM;
1769
1770         /* __khugepaged_exit() must not run from under us */
1771         VM_BUG_ON(khugepaged_test_exit(mm));
1772         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1773                 free_mm_slot(mm_slot);
1774                 return 0;
1775         }
1776
1777         spin_lock(&khugepaged_mm_lock);
1778         insert_to_mm_slots_hash(mm, mm_slot);
1779         /*
1780          * Insert just behind the scanning cursor, to let the area settle
1781          * down a little.
1782          */
1783         wakeup = list_empty(&khugepaged_scan.mm_head);
1784         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1785         spin_unlock(&khugepaged_mm_lock);
1786
1787         atomic_inc(&mm->mm_count);
1788         if (wakeup)
1789                 wake_up_interruptible(&khugepaged_wait);
1790
1791         return 0;
1792 }
1793
1794 int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
1795 {
1796         unsigned long hstart, hend;
1797         if (!vma->anon_vma)
1798                 /*
1799                  * Not yet faulted in so we will register later in the
1800                  * page fault if needed.
1801                  */
1802                 return 0;
1803         if (vma->vm_ops)
1804                 /* khugepaged not yet working on file or special mappings */
1805                 return 0;
1806         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
1807         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1808         hend = vma->vm_end & HPAGE_PMD_MASK;
1809         if (hstart < hend)
1810                 return khugepaged_enter(vma);
1811         return 0;
1812 }
1813
1814 void __khugepaged_exit(struct mm_struct *mm)
1815 {
1816         struct mm_slot *mm_slot;
1817         int free = 0;
1818
1819         spin_lock(&khugepaged_mm_lock);
1820         mm_slot = get_mm_slot(mm);
1821         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1822                 hlist_del(&mm_slot->hash);
1823                 list_del(&mm_slot->mm_node);
1824                 free = 1;
1825         }
1826         spin_unlock(&khugepaged_mm_lock);
1827
1828         if (free) {
1829                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1830                 free_mm_slot(mm_slot);
1831                 mmdrop(mm);
1832         } else if (mm_slot) {
1833                 /*
1834                  * This is required to serialize against
1835                  * khugepaged_test_exit() (which is guaranteed to run
1836                  * under mmap sem read mode). Stop here (after we
1837                  * return all pagetables will be destroyed) until
1838                  * khugepaged has finished working on the pagetables
1839                  * under the mmap_sem.
1840                  */
1841                 down_write(&mm->mmap_sem);
1842                 up_write(&mm->mmap_sem);
1843         }
1844 }
1845
1846 static void release_pte_page(struct page *page)
1847 {
1848         /* 0 stands for page_is_file_cache(page) == false */
1849         dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
1850         unlock_page(page);
1851         putback_lru_page(page);
1852 }
1853
1854 static void release_pte_pages(pte_t *pte, pte_t *_pte)
1855 {
1856         while (--_pte >= pte) {
1857                 pte_t pteval = *_pte;
1858                 if (!pte_none(pteval))
1859                         release_pte_page(pte_page(pteval));
1860         }
1861 }
1862
1863 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
1864                                         unsigned long address,
1865                                         pte_t *pte)
1866 {
1867         struct page *page;
1868         pte_t *_pte;
1869         int referenced = 0, none = 0;
1870         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
1871              _pte++, address += PAGE_SIZE) {
1872                 pte_t pteval = *_pte;
1873                 if (pte_none(pteval)) {
1874                         if (++none <= khugepaged_max_ptes_none)
1875                                 continue;
1876                         else
1877                                 goto out;
1878                 }
1879                 if (!pte_present(pteval) || !pte_write(pteval))
1880                         goto out;
1881                 page = vm_normal_page(vma, address, pteval);
1882                 if (unlikely(!page))
1883                         goto out;
1884
1885                 VM_BUG_ON(PageCompound(page));
1886                 BUG_ON(!PageAnon(page));
1887                 VM_BUG_ON(!PageSwapBacked(page));
1888
1889                 /* cannot use mapcount: can't collapse if there's a gup pin */
1890                 if (page_count(page) != 1)
1891                         goto out;
1892                 /*
1893                  * We can do it before isolate_lru_page because the
1894                  * page can't be freed from under us. NOTE: PG_lock
1895                  * is needed to serialize against split_huge_page
1896                  * when invoked from the VM.
1897                  */
1898                 if (!trylock_page(page))
1899                         goto out;
1900                 /*
1901                  * Isolate the page to avoid collapsing an hugepage
1902                  * currently in use by the VM.
1903                  */
1904                 if (isolate_lru_page(page)) {
1905                         unlock_page(page);
1906                         goto out;
1907                 }
1908                 /* 0 stands for page_is_file_cache(page) == false */
1909                 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
1910                 VM_BUG_ON(!PageLocked(page));
1911                 VM_BUG_ON(PageLRU(page));
1912
1913                 /* If there is no mapped pte young don't collapse the page */
1914                 if (pte_young(pteval) || PageReferenced(page) ||
1915                     mmu_notifier_test_young(vma->vm_mm, address))
1916                         referenced = 1;
1917         }
1918         if (likely(referenced))
1919                 return 1;
1920 out:
1921         release_pte_pages(pte, _pte);
1922         return 0;
1923 }
1924
1925 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
1926                                       struct vm_area_struct *vma,
1927                                       unsigned long address,
1928                                       spinlock_t *ptl)
1929 {
1930         pte_t *_pte;
1931         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
1932                 pte_t pteval = *_pte;
1933                 struct page *src_page;
1934
1935                 if (pte_none(pteval)) {
1936                         clear_user_highpage(page, address);
1937                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
1938                 } else {
1939                         src_page = pte_page(pteval);
1940                         copy_user_highpage(page, src_page, address, vma);
1941                         VM_BUG_ON(page_mapcount(src_page) != 1);
1942                         release_pte_page(src_page);
1943                         /*
1944                          * ptl mostly unnecessary, but preempt has to
1945                          * be disabled to update the per-cpu stats
1946                          * inside page_remove_rmap().
1947                          */
1948                         spin_lock(ptl);
1949                         /*
1950                          * paravirt calls inside pte_clear here are
1951                          * superfluous.
1952                          */
1953                         pte_clear(vma->vm_mm, address, _pte);
1954                         page_remove_rmap(src_page);
1955                         spin_unlock(ptl);
1956                         free_page_and_swap_cache(src_page);
1957                 }
1958
1959                 address += PAGE_SIZE;
1960                 page++;
1961         }
1962 }
1963
1964 static void khugepaged_alloc_sleep(void)
1965 {
1966         wait_event_freezable_timeout(khugepaged_wait, false,
1967                         msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
1968 }
1969
1970 #ifdef CONFIG_NUMA
1971 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
1972 {
1973         if (IS_ERR(*hpage)) {
1974                 if (!*wait)
1975                         return false;
1976
1977                 *wait = false;
1978                 *hpage = NULL;
1979                 khugepaged_alloc_sleep();
1980         } else if (*hpage) {
1981                 put_page(*hpage);
1982                 *hpage = NULL;
1983         }
1984
1985         return true;
1986 }
1987
1988 static struct page
1989 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
1990                        struct vm_area_struct *vma, unsigned long address,
1991                        int node)
1992 {
1993         VM_BUG_ON(*hpage);
1994         /*
1995          * Allocate the page while the vma is still valid and under
1996          * the mmap_sem read mode so there is no memory allocation
1997          * later when we take the mmap_sem in write mode. This is more
1998          * friendly behavior (OTOH it may actually hide bugs) to
1999          * filesystems in userland with daemons allocating memory in
2000          * the userland I/O paths.  Allocating memory with the
2001          * mmap_sem in read mode is good idea also to allow greater
2002          * scalability.
2003          */
2004         *hpage  = alloc_hugepage_vma(khugepaged_defrag(), vma, address,
2005                                       node, __GFP_OTHER_NODE);
2006
2007         /*
2008          * After allocating the hugepage, release the mmap_sem read lock in
2009          * preparation for taking it in write mode.
2010          */
2011         up_read(&mm->mmap_sem);
2012         if (unlikely(!*hpage)) {
2013                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2014                 *hpage = ERR_PTR(-ENOMEM);
2015                 return NULL;
2016         }
2017
2018         count_vm_event(THP_COLLAPSE_ALLOC);
2019         return *hpage;
2020 }
2021 #else
2022 static struct page *khugepaged_alloc_hugepage(bool *wait)
2023 {
2024         struct page *hpage;
2025
2026         do {
2027                 hpage = alloc_hugepage(khugepaged_defrag());
2028                 if (!hpage) {
2029                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2030                         if (!*wait)
2031                                 return NULL;
2032
2033                         *wait = false;
2034                         khugepaged_alloc_sleep();
2035                 } else
2036                         count_vm_event(THP_COLLAPSE_ALLOC);
2037         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2038
2039         return hpage;
2040 }
2041
2042 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2043 {
2044         if (!*hpage)
2045                 *hpage = khugepaged_alloc_hugepage(wait);
2046
2047         if (unlikely(!*hpage))
2048                 return false;
2049
2050         return true;
2051 }
2052
2053 static struct page
2054 *khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2055                        struct vm_area_struct *vma, unsigned long address,
2056                        int node)
2057 {
2058         up_read(&mm->mmap_sem);
2059         VM_BUG_ON(!*hpage);
2060         return  *hpage;
2061 }
2062 #endif
2063
2064 static bool hugepage_vma_check(struct vm_area_struct *vma)
2065 {
2066         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2067             (vma->vm_flags & VM_NOHUGEPAGE))
2068                 return false;
2069
2070         if (!vma->anon_vma || vma->vm_ops)
2071                 return false;
2072         if (is_vma_temporary_stack(vma))
2073                 return false;
2074         VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2075         return true;
2076 }
2077
2078 static void collapse_huge_page(struct mm_struct *mm,
2079                                    unsigned long address,
2080                                    struct page **hpage,
2081                                    struct vm_area_struct *vma,
2082                                    int node)
2083 {
2084         pmd_t *pmd, _pmd;
2085         pte_t *pte;
2086         pgtable_t pgtable;
2087         struct page *new_page;
2088         spinlock_t *ptl;
2089         int isolated;
2090         unsigned long hstart, hend;
2091         unsigned long mmun_start;       /* For mmu_notifiers */
2092         unsigned long mmun_end;         /* For mmu_notifiers */
2093
2094         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2095
2096         /* release the mmap_sem read lock. */
2097         new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2098         if (!new_page)
2099                 return;
2100
2101         if (unlikely(mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL)))
2102                 return;
2103
2104         /*
2105          * Prevent all access to pagetables with the exception of
2106          * gup_fast later hanlded by the ptep_clear_flush and the VM
2107          * handled by the anon_vma lock + PG_lock.
2108          */
2109         down_write(&mm->mmap_sem);
2110         if (unlikely(khugepaged_test_exit(mm)))
2111                 goto out;
2112
2113         vma = find_vma(mm, address);
2114         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2115         hend = vma->vm_end & HPAGE_PMD_MASK;
2116         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2117                 goto out;
2118         if (!hugepage_vma_check(vma))
2119                 goto out;
2120         pmd = mm_find_pmd(mm, address);
2121         if (!pmd)
2122                 goto out;
2123         if (pmd_trans_huge(*pmd))
2124                 goto out;
2125
2126         anon_vma_lock(vma->anon_vma);
2127
2128         pte = pte_offset_map(pmd, address);
2129         ptl = pte_lockptr(mm, pmd);
2130
2131         mmun_start = address;
2132         mmun_end   = address + HPAGE_PMD_SIZE;
2133         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2134         spin_lock(&mm->page_table_lock); /* probably unnecessary */
2135         /*
2136          * After this gup_fast can't run anymore. This also removes
2137          * any huge TLB entry from the CPU so we won't allow
2138          * huge and small TLB entries for the same virtual address
2139          * to avoid the risk of CPU bugs in that area.
2140          */
2141         _pmd = pmdp_clear_flush(vma, address, pmd);
2142         spin_unlock(&mm->page_table_lock);
2143         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2144
2145         spin_lock(ptl);
2146         isolated = __collapse_huge_page_isolate(vma, address, pte);
2147         spin_unlock(ptl);
2148
2149         if (unlikely(!isolated)) {
2150                 pte_unmap(pte);
2151                 spin_lock(&mm->page_table_lock);
2152                 BUG_ON(!pmd_none(*pmd));
2153                 set_pmd_at(mm, address, pmd, _pmd);
2154                 spin_unlock(&mm->page_table_lock);
2155                 anon_vma_unlock(vma->anon_vma);
2156                 goto out;
2157         }
2158
2159         /*
2160          * All pages are isolated and locked so anon_vma rmap
2161          * can't run anymore.
2162          */
2163         anon_vma_unlock(vma->anon_vma);
2164
2165         __collapse_huge_page_copy(pte, new_page, vma, address, ptl);
2166         pte_unmap(pte);
2167         __SetPageUptodate(new_page);
2168         pgtable = pmd_pgtable(_pmd);
2169
2170         _pmd = mk_huge_pmd(new_page, vma);
2171
2172         /*
2173          * spin_lock() below is not the equivalent of smp_wmb(), so
2174          * this is needed to avoid the copy_huge_page writes to become
2175          * visible after the set_pmd_at() write.
2176          */
2177         smp_wmb();
2178
2179         spin_lock(&mm->page_table_lock);
2180         BUG_ON(!pmd_none(*pmd));
2181         page_add_new_anon_rmap(new_page, vma, address);
2182         set_pmd_at(mm, address, pmd, _pmd);
2183         update_mmu_cache_pmd(vma, address, pmd);
2184         pgtable_trans_huge_deposit(mm, pgtable);
2185         spin_unlock(&mm->page_table_lock);
2186
2187         *hpage = NULL;
2188
2189         khugepaged_pages_collapsed++;
2190 out_up_write:
2191         up_write(&mm->mmap_sem);
2192         return;
2193
2194 out:
2195         mem_cgroup_uncharge_page(new_page);
2196         goto out_up_write;
2197 }
2198
2199 static int khugepaged_scan_pmd(struct mm_struct *mm,
2200                                struct vm_area_struct *vma,
2201                                unsigned long address,
2202                                struct page **hpage)
2203 {
2204         pmd_t *pmd;
2205         pte_t *pte, *_pte;
2206         int ret = 0, referenced = 0, none = 0;
2207         struct page *page;
2208         unsigned long _address;
2209         spinlock_t *ptl;
2210         int node = -1;
2211
2212         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2213
2214         pmd = mm_find_pmd(mm, address);
2215         if (!pmd)
2216                 goto out;
2217         if (pmd_trans_huge(*pmd))
2218                 goto out;
2219
2220         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2221         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2222              _pte++, _address += PAGE_SIZE) {
2223                 pte_t pteval = *_pte;
2224                 if (pte_none(pteval)) {
2225                         if (++none <= khugepaged_max_ptes_none)
2226                                 continue;
2227                         else
2228                                 goto out_unmap;
2229                 }
2230                 if (!pte_present(pteval) || !pte_write(pteval))
2231                         goto out_unmap;
2232                 page = vm_normal_page(vma, _address, pteval);
2233                 if (unlikely(!page))
2234                         goto out_unmap;
2235                 /*
2236                  * Chose the node of the first page. This could
2237                  * be more sophisticated and look at more pages,
2238                  * but isn't for now.
2239                  */
2240                 if (node == -1)
2241                         node = page_to_nid(page);
2242                 VM_BUG_ON(PageCompound(page));
2243                 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2244                         goto out_unmap;
2245                 /* cannot use mapcount: can't collapse if there's a gup pin */
2246                 if (page_count(page) != 1)
2247                         goto out_unmap;
2248                 if (pte_young(pteval) || PageReferenced(page) ||
2249                     mmu_notifier_test_young(vma->vm_mm, address))
2250                         referenced = 1;
2251         }
2252         if (referenced)
2253                 ret = 1;
2254 out_unmap:
2255         pte_unmap_unlock(pte, ptl);
2256         if (ret)
2257                 /* collapse_huge_page will return with the mmap_sem released */
2258                 collapse_huge_page(mm, address, hpage, vma, node);
2259 out:
2260         return ret;
2261 }
2262
2263 static void collect_mm_slot(struct mm_slot *mm_slot)
2264 {
2265         struct mm_struct *mm = mm_slot->mm;
2266
2267         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2268
2269         if (khugepaged_test_exit(mm)) {
2270                 /* free mm_slot */
2271                 hlist_del(&mm_slot->hash);
2272                 list_del(&mm_slot->mm_node);
2273
2274                 /*
2275                  * Not strictly needed because the mm exited already.
2276                  *
2277                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2278                  */
2279
2280                 /* khugepaged_mm_lock actually not necessary for the below */
2281                 free_mm_slot(mm_slot);
2282                 mmdrop(mm);
2283         }
2284 }
2285
2286 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2287                                             struct page **hpage)
2288         __releases(&khugepaged_mm_lock)
2289         __acquires(&khugepaged_mm_lock)
2290 {
2291         struct mm_slot *mm_slot;
2292         struct mm_struct *mm;
2293         struct vm_area_struct *vma;
2294         int progress = 0;
2295
2296         VM_BUG_ON(!pages);
2297         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2298
2299         if (khugepaged_scan.mm_slot)
2300                 mm_slot = khugepaged_scan.mm_slot;
2301         else {
2302                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2303                                      struct mm_slot, mm_node);
2304                 khugepaged_scan.address = 0;
2305                 khugepaged_scan.mm_slot = mm_slot;
2306         }
2307         spin_unlock(&khugepaged_mm_lock);
2308
2309         mm = mm_slot->mm;
2310         down_read(&mm->mmap_sem);
2311         if (unlikely(khugepaged_test_exit(mm)))
2312                 vma = NULL;
2313         else
2314                 vma = find_vma(mm, khugepaged_scan.address);
2315
2316         progress++;
2317         for (; vma; vma = vma->vm_next) {
2318                 unsigned long hstart, hend;
2319
2320                 cond_resched();
2321                 if (unlikely(khugepaged_test_exit(mm))) {
2322                         progress++;
2323                         break;
2324                 }
2325                 if (!hugepage_vma_check(vma)) {
2326 skip:
2327                         progress++;
2328                         continue;
2329                 }
2330                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2331                 hend = vma->vm_end & HPAGE_PMD_MASK;
2332                 if (hstart >= hend)
2333                         goto skip;
2334                 if (khugepaged_scan.address > hend)
2335                         goto skip;
2336                 if (khugepaged_scan.address < hstart)
2337                         khugepaged_scan.address = hstart;
2338                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2339
2340                 while (khugepaged_scan.address < hend) {
2341                         int ret;
2342                         cond_resched();
2343                         if (unlikely(khugepaged_test_exit(mm)))
2344                                 goto breakouterloop;
2345
2346                         VM_BUG_ON(khugepaged_scan.address < hstart ||
2347                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
2348                                   hend);
2349                         ret = khugepaged_scan_pmd(mm, vma,
2350                                                   khugepaged_scan.address,
2351                                                   hpage);
2352                         /* move to next address */
2353                         khugepaged_scan.address += HPAGE_PMD_SIZE;
2354                         progress += HPAGE_PMD_NR;
2355                         if (ret)
2356                                 /* we released mmap_sem so break loop */
2357                                 goto breakouterloop_mmap_sem;
2358                         if (progress >= pages)
2359                                 goto breakouterloop;
2360                 }
2361         }
2362 breakouterloop:
2363         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2364 breakouterloop_mmap_sem:
2365
2366         spin_lock(&khugepaged_mm_lock);
2367         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2368         /*
2369          * Release the current mm_slot if this mm is about to die, or
2370          * if we scanned all vmas of this mm.
2371          */
2372         if (khugepaged_test_exit(mm) || !vma) {
2373                 /*
2374                  * Make sure that if mm_users is reaching zero while
2375                  * khugepaged runs here, khugepaged_exit will find
2376                  * mm_slot not pointing to the exiting mm.
2377                  */
2378                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2379                         khugepaged_scan.mm_slot = list_entry(
2380                                 mm_slot->mm_node.next,
2381                                 struct mm_slot, mm_node);
2382                         khugepaged_scan.address = 0;
2383                 } else {
2384                         khugepaged_scan.mm_slot = NULL;
2385                         khugepaged_full_scans++;
2386                 }
2387
2388                 collect_mm_slot(mm_slot);
2389         }
2390
2391         return progress;
2392 }
2393
2394 static int khugepaged_has_work(void)
2395 {
2396         return !list_empty(&khugepaged_scan.mm_head) &&
2397                 khugepaged_enabled();
2398 }
2399
2400 static int khugepaged_wait_event(void)
2401 {
2402         return !list_empty(&khugepaged_scan.mm_head) ||
2403                 kthread_should_stop();
2404 }
2405
2406 static void khugepaged_do_scan(void)
2407 {
2408         struct page *hpage = NULL;
2409         unsigned int progress = 0, pass_through_head = 0;
2410         unsigned int pages = khugepaged_pages_to_scan;
2411         bool wait = true;
2412
2413         barrier(); /* write khugepaged_pages_to_scan to local stack */
2414
2415         while (progress < pages) {
2416                 if (!khugepaged_prealloc_page(&hpage, &wait))
2417                         break;
2418
2419                 cond_resched();
2420
2421                 if (unlikely(kthread_should_stop() || freezing(current)))
2422                         break;
2423
2424                 spin_lock(&khugepaged_mm_lock);
2425                 if (!khugepaged_scan.mm_slot)
2426                         pass_through_head++;
2427                 if (khugepaged_has_work() &&
2428                     pass_through_head < 2)
2429                         progress += khugepaged_scan_mm_slot(pages - progress,
2430                                                             &hpage);
2431                 else
2432                         progress = pages;
2433                 spin_unlock(&khugepaged_mm_lock);
2434         }
2435
2436         if (!IS_ERR_OR_NULL(hpage))
2437                 put_page(hpage);
2438 }
2439
2440 static void khugepaged_wait_work(void)
2441 {
2442         try_to_freeze();
2443
2444         if (khugepaged_has_work()) {
2445                 if (!khugepaged_scan_sleep_millisecs)
2446                         return;
2447
2448                 wait_event_freezable_timeout(khugepaged_wait,
2449                                              kthread_should_stop(),
2450                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2451                 return;
2452         }
2453
2454         if (khugepaged_enabled())
2455                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2456 }
2457
2458 static int khugepaged(void *none)
2459 {
2460         struct mm_slot *mm_slot;
2461
2462         set_freezable();
2463         set_user_nice(current, 19);
2464
2465         while (!kthread_should_stop()) {
2466                 khugepaged_do_scan();
2467                 khugepaged_wait_work();
2468         }
2469
2470         spin_lock(&khugepaged_mm_lock);
2471         mm_slot = khugepaged_scan.mm_slot;
2472         khugepaged_scan.mm_slot = NULL;
2473         if (mm_slot)
2474                 collect_mm_slot(mm_slot);
2475         spin_unlock(&khugepaged_mm_lock);
2476         return 0;
2477 }
2478
2479 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2480                 unsigned long haddr, pmd_t *pmd)
2481 {
2482         struct mm_struct *mm = vma->vm_mm;
2483         pgtable_t pgtable;
2484         pmd_t _pmd;
2485         int i;
2486
2487         pmdp_clear_flush(vma, haddr, pmd);
2488         /* leave pmd empty until pte is filled */
2489
2490         pgtable = pgtable_trans_huge_withdraw(mm);
2491         pmd_populate(mm, &_pmd, pgtable);
2492
2493         for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2494                 pte_t *pte, entry;
2495                 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2496                 entry = pte_mkspecial(entry);
2497                 pte = pte_offset_map(&_pmd, haddr);
2498                 VM_BUG_ON(!pte_none(*pte));
2499                 set_pte_at(mm, haddr, pte, entry);
2500                 pte_unmap(pte);
2501         }
2502         smp_wmb(); /* make pte visible before pmd */
2503         pmd_populate(mm, pmd, pgtable);
2504 }
2505
2506 void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2507                 pmd_t *pmd)
2508 {
2509         struct page *page;
2510         struct mm_struct *mm = vma->vm_mm;
2511         unsigned long haddr = address & HPAGE_PMD_MASK;
2512         unsigned long mmun_start;       /* For mmu_notifiers */
2513         unsigned long mmun_end;         /* For mmu_notifiers */
2514
2515         BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2516
2517         mmun_start = haddr;
2518         mmun_end   = haddr + HPAGE_PMD_SIZE;
2519         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2520         spin_lock(&mm->page_table_lock);
2521         if (unlikely(!pmd_trans_huge(*pmd))) {
2522                 spin_unlock(&mm->page_table_lock);
2523                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2524                 return;
2525         }
2526         if (is_huge_zero_pmd(*pmd)) {
2527                 __split_huge_zero_page_pmd(vma, haddr, pmd);
2528                 spin_unlock(&mm->page_table_lock);
2529                 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2530                 return;
2531         }
2532         page = pmd_page(*pmd);
2533         VM_BUG_ON(!page_count(page));
2534         get_page(page);
2535         spin_unlock(&mm->page_table_lock);
2536         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2537
2538         split_huge_page(page);
2539
2540         put_page(page);
2541         BUG_ON(pmd_trans_huge(*pmd));
2542 }
2543
2544 void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2545                 pmd_t *pmd)
2546 {
2547         struct vm_area_struct *vma;
2548
2549         vma = find_vma(mm, address);
2550         BUG_ON(vma == NULL);
2551         split_huge_page_pmd(vma, address, pmd);
2552 }
2553
2554 static void split_huge_page_address(struct mm_struct *mm,
2555                                     unsigned long address)
2556 {
2557         pmd_t *pmd;
2558
2559         VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2560
2561         pmd = mm_find_pmd(mm, address);
2562         if (!pmd)
2563                 return;
2564         /*
2565          * Caller holds the mmap_sem write mode, so a huge pmd cannot
2566          * materialize from under us.
2567          */
2568         split_huge_page_pmd_mm(mm, address, pmd);
2569 }
2570
2571 void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2572                              unsigned long start,
2573                              unsigned long end,
2574                              long adjust_next)
2575 {
2576         /*
2577          * If the new start address isn't hpage aligned and it could
2578          * previously contain an hugepage: check if we need to split
2579          * an huge pmd.
2580          */
2581         if (start & ~HPAGE_PMD_MASK &&
2582             (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2583             (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2584                 split_huge_page_address(vma->vm_mm, start);
2585
2586         /*
2587          * If the new end address isn't hpage aligned and it could
2588          * previously contain an hugepage: check if we need to split
2589          * an huge pmd.
2590          */
2591         if (end & ~HPAGE_PMD_MASK &&
2592             (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2593             (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2594                 split_huge_page_address(vma->vm_mm, end);
2595
2596         /*
2597          * If we're also updating the vma->vm_next->vm_start, if the new
2598          * vm_next->vm_start isn't page aligned and it could previously
2599          * contain an hugepage: check if we need to split an huge pmd.
2600          */
2601         if (adjust_next > 0) {
2602                 struct vm_area_struct *next = vma->vm_next;
2603                 unsigned long nstart = next->vm_start;
2604                 nstart += adjust_next << PAGE_SHIFT;
2605                 if (nstart & ~HPAGE_PMD_MASK &&
2606                     (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2607                     (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2608                         split_huge_page_address(next->vm_mm, nstart);
2609         }
2610 }